Computers, Environment and Urban Systems

31 (2007) 393–408
www.elsevier.com/locate/compenvurbsys

Simultaneous storage of primal and dual

three-dimensional subdivisions
a,* b
Hugo Ledoux , Christopher M. Gold
a
Delft University of Technology, (OTB, Section GIS Technology), Jaffalaan 9, 2628BX Delft, The Netherlands
b
GIS Research Centre, School of Computing, University of Glamorgan, Pontypridd CF37 1DL, Wales, UK

Received 9 September 2005; accepted 14 March 2006

Abstract

We propose a new general-purpose data structure useful for a variety of three-dimensional appli-
cations. The data structure has the characteristic of storing simultaneously the primal and dual sub-
divisions of a three-dimensional manifold. We argue in this paper that storing both subdivisions, for
instance the Voronoi diagram and the Delaunay tetrahedralization, can be beneficial for many appli-
cation domains, notably for the modelling of datasets in geosciences or for representing boundaries
of real-world features. Our structure is an extension of the well-known quad-edge data structure used
for representing two-dimensional manifolds. We describe the basic properties of this augmented
quad-edge structure, along with the navigation operators, and we also demonstrate its usefulness
with some examples of applications.
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Data structure; Three-dimensional modelling; Duality; Voronoi diagram; Delaunay tetrahedra-
lization

1. Introduction

The topological data structures commonly found in geographical information systems

(GIS) or other modelling systems have been used – with success – for many years to model
two-dimensional (2D) objects and surfaces (e.g. a digital terrain model where the elevation

*
Corresponding author. Tel.: +31 15 2786114; fax: +31 15 2782745.
E-mail addresses: [email protected] (H. Ledoux), [email protected] (C.M. Gold).

0198-9715/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compenvurbsys.2006.03.003
394 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

is treated as an attribute). The TIGER (Boudriault, 1987), DCEL (Muller & Preparata,
1978) and ARC/INFO (Morehouse, 1985) formats are all examples of such structures.
Although many application domains require more than that, data structures to model
three-dimensional (3D) solid objects (not only their surface, but also their interior) are still
not well-studied and can even be considered to be in their infancy. Examples of domains
where 3D data structures are required include computer graphics/games, CAD systems,
medical imaging, and the modelling of cities or geoscientific datasets.
In this paper, we propose a new general-purpose data structure useful for a variety of
applications where 3D solid objects are involved. To represent such objects in computers,
the space they cover has to be decomposed, or subdivided, into finite parts, i.e., into volume
elements that we refer to as cells. For the purpose of this paper, we are particularly inter-
ested in data structures that permit us to store not only one subdivision of a 3D object, but
simultaneously both that subdivision (which is refer to as the primal) and another related
subdivision called the dual. The terms ‘primal’ and ‘dual’ refer to the concept of duality
in mathematics, as explained in Section 3. In brief, it means that two subdivisions are
inter-connected: they represent the same thing, just from a different point of view. The most
known and used example of primal and dual subdivisions is the duality that exists in any
dimensions between the Voronoi diagram (VD) and the Delaunay triangulation (DT).
Our claim in this paper is that storing both the primal and the dual subdivisions can be
beneficial for a broad range of applications, and that doing so offers a more powerful and
flexible solution than the other known 3D data structures (see Section 2 for a review). As
explained in Section 5, keeping both subdivisions can optimise certain spatial analysis
operations in GIS, and permit new applications involving the volume and the boundary
of an object. Also, it offers a flexible solution, because attributes for any elements of both
subdivisions can be stored, which is of considerable importance for the modelling of real-
world features. On the other hand, keeping two subdivisions at the same time obviously
has drawbacks: it increases the requirements for storage, and the construction/modifica-
tion algorithms will be slower. We nevertheless believe that in many real-world applica-
tions the major constraint is not the speed of construction of the topological models of
a large number of points or the space needed in memory, but rather the ability to interac-
tively construct, edit (by deleting or moving certain points) and query (interpolation,
extraction of implicit surfaces, etc.) the desired model. See Anselin (1999) and Gold
(1993) for discussions about the advantages of interactive modelling.
We present in Section 4 a new data structure to store and manipulate simultaneously
the primal and dual subdivisions of a 3D object. The structure, called the augmented
quad-edge (AQE), is based on the quad-edge structure (Guibas & Stolfi, 1985) that has been
used for many years to represent the primal and dual subdivisions of 2D objects. The basic
properties, the construction and navigation operators, and the storage costs of the AQE
are also presented in Section 4. We believe the primary feature of the data structure to
be ease of use in managing and navigating 3D objects. While it needs somewhat more stor-
age than other structures, it is computationally efficient, requiring no searching operations
in order to move from cell to cell.

2. Related work

Before discussing the alternative data structures, we need to introduce some concepts
and definitions. When dealing with d-dimensional objects, an important concept is that
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 395

Fig. 1. (a) The neighbourhood of every point in a 2-manifold is equivalent to a plane. (b) The neighbourhood of
every point in a 3-manifold is equivalent to a 3D space. (c) A non-manifold object. The neighbourhood of the
vertex where the two tetrahedra, the triangular face and the edge meet does not resemble a 3D space, i.e., a
coordinate system can not be defined locally.

of a manifold. Informally, a d-manifold is an object for which the neighbourhood of every

one of its points resembles a d-dimensional space. A 2-manifold is therefore a surface
which, it should be pointed out, can be embedded in 3D space (see Fig. 1a); an obvious
example is the surface of the Earth, for which near to every point the environment is equiv-
alent to a plane. An example of a 3-manifold is the entire Earth (its interior) because the
neighbourhood of every point is equivalent to a 3D region (see Fig. 1b). Closely related to
the concept of a subdivision of a 3-manifold is that of a 3D cell complex. A cell complex is
formed by a collection of k-cells (where 0 < k 6 d, such that a 0-cell is a vertex, a 1-cell an
edge, a 2-cell a face and a 3-cell a polyhedron) such that no k-cell punctures the interior of
any other k-cell. In the following, unless explicitly stated, a cell is always assumed to be a
3-cell. Note that a 3D cell complex is the same as the tessellation of 3D space into non-
overlapping polyhedra.
Data structures to represent 3D objects have been developed in many different fields,
and we would classify the most common 3D spatial data structures as follows. Note that
this is not a complete review, our purpose is to introduce some of the methods used in
other disciplines.

Constructive solid geometry (CSG): Solid objects are represented as Boolean combina-
tions (union, intersection and difference) of simpler objects, which are typically of sim-
ple shapes such as cylinders, cones, spheres or pyramids (see Requicha (1982) for more
details). CSG is mostly used in CAD systems to represent complex models of man-made
features (e.g. aircrafts or engines).
Boundary representation (b-rep): One 3D object – it must be a 2-manifold – is repre-
sented by its boundary surface, using primitive objects (vertices, edges and faces).
Boundary modelling is more flexible than CSG because irregular objects can be mod-
elled, but it requires more work as the information about the connectivity of the faces
needs to be valid at all times. One example of the use of such a structure is the modelling
of cities in a GIS where each building is represented individually by a b-rep (see Zlata-
nova, Abdul Rahman, & Shi (2004) for a review of the topological models used).
Regular decomposition: The space covered by an object is arbitrarily decomposed into
cells of regular shape (usually cubes in 3D). The term ‘voxel’ is used as the 3D
396 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

equivalent of pixel. This structure, or closely related hierarchical decompositions

(Samet, 1984), have been used in many application domains to represent continuous
phenomena.
Irregular decomposition: The object is decomposed into cells of different size and/or
shape. The cells can all have the same shape (e.g. tetrahedra), or be arbitrary polyhedra.
It is equivalent to a subdivision or a tessellation of a 3-manifold.
Non-manifold structures: These are used to represent non-manifold objects, i.e., objects
for which the neighbourhood of a point does not resemble a 3D region (see Fig. 1c). See
De Floriani and Hui (2003) for one implementation of such a structure.

Let us now describe in more detail some of the most relevant data structures. To rep-
resent 2-manifolds, only the edges and vertices of a subdivision are usually stored, and for
each edge the connected edges are stored; this permits us to store implicitly the faces of the
boundary. The most popular structures are the half-edge (Mäntylä, 1988), the DCEL
(Muller & Preparata, 1978) and the winged-edge (Baumgart, 1975). Guibas and Stolfi
(1985) also propose the quad-edge, which stores simultaneously any primal and dual sub-
divisions of a 2-manifold. The structure, which is formally described in Section 4.1, comes
with an algebra for the construction and the navigation of the subdivisions.
Many data structures to store arbitrary 3D subdivisions are extensions of work done in
2D. Examples are the half-face of Lopes and Tavares (1997) (extension of the half-edge)
and the Generalized Maps (G-Maps) (Bertrand, Dufourd, Françon, & Lienhardt, 1993;
Lienhardt, 1994). The latter structure, which is actually valid to represent a broader class
of objects called cellular quasi-manifolds, can be seen as a generalization to d-manifolds of
boundary models as each k-cell (where 0 < k 6 d) is recursively decomposed into cells of
lower dimensionality, and the topological relationships between adjacent k-cells are kept.
The resulting structure is thus very space-consuming, as each lower-dimensionality cell
must be stored in cells of higher dimensionality: a vertex shared by five edges is for
instance stored in each of the five edges. It is nevertheless being used in a commercial sys-
tem for the modelling of geoscientific data.1 Although G-Maps store only one subdivision,
it offers efficient functions to extract all the elements of the dual subdivision. One of the
few structures permitting the simultaneous preservation of primal and dual subdivisions
of 3-manifolds is the facet-edge structure of Dobkin and Laszlo (1989). As its name
implies, the atom is a pair formed of a face and an edge, and the next facet-edge pairs
(one in each direction) around the face and the edge (incident faces to the edge) are stored.
Unlike G-Maps, a face shared by two cells needs only be stored once. The facet-edge struc-
ture comes with a set of operations to modify cells and to navigate within both subdivi-
sions. Its generality makes navigation within a single cell impossible, and hence the
authors suggest storing extra information for each edge and using the quad-edge opera-
tors. Also, unlike the quad-edge that is being used in many implementations of the 2D
VD/DT, the facet-edge has been found difficult to implement in practice and, to our
knowledge, has not been used in ‘real projects’.
As Aurenhammer (1991) demonstrates, the VD and the DT are fundamental concepts
in computer science because they are useful in a wide variety of applications, and once they
are constructed they permit us to solve many geometric problems. In fact, they are such

1
The gOcad system: <www.gocad.org>.
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 397

important concepts that it is worth mentioning data structures to store these in particular.
In practice, most of the algorithms and implementations available to construct the 3D
VD/DT store only the DT and perform their topological operations on tetrahedra, and
if needed the VD is extracted afterwards. This is usually justified by the fact that the
knowledge of either structure implies the knowledge of the other, and also because man-
aging and storing simplices – having a fixed number of vertices and adjacent cells – over
arbitrary polyhedra simplifies the topological operations and permits the design of simpler
data structures. The resulting implementation will therefore be more efficient, both in
terms of speed and memory. The simplest data structure considers the tetrahedron as
being its atom and stores each tetrahedron with four pointers to its vertices and four point-
ers to its adjacent tetrahedra. This is a very space-efficient structure that yields fast imple-
mentations, and it is therefore the structure of choice in many projects. The CGAL library
notably uses it, but has added to each vertex a pointer to one of its incident tetrahedra
to speed up the extraction of the Voronoi cells (Boissonnat, Devillers, Pion, Teillaud, &
Yvinec, 2002). It is also possible to store the tetrahedra implicitly by considering a data
structure where the atom is a triangular face having three pointers to its vertices and six
pointers to its adjacent faces, as described by Shewchuk (1997). However, storing only
the DT has major drawbacks if one wants to work with the VD. It is for example difficult
to assign attributes to Voronoi vertices or faces, and moreover the computation of the VD
is an expensive operation, as explained in Section 5.2.

3. Duality in three-dimensional space

Duality can have many different meanings in mathematics, but it always refers to the
translation or mapping in a one-to-one fashion of concepts or structures. We use it here
in the sense of the dual graph of a given graph. Let G be a planar graph, as illustrated
in Fig. 2a. Note that G can both be seen as a subdivision of a 2-manifold, or as a 2D cell
complex. The dual graph Gw has a vertex for each polygon in G, and the vertices in Gw are
linked by an edge if and only if the two corresponding dual polygons in G are adjacent (in
Fig. 2b, Gw is represented with dashed lines). Notice also that each polygon in Gw corres-
ponds to a vertex in G. Consider now the polygonal map in Fig. 2c, where each polygon (a
map object) has certain attributes, and where the boundaries between two map objects
also have some sort of attributes, e.g. the boundary form or the flow. Notice that the
boundaries do not characterise per se any of the objects, but rather the adjacency relation-
ships that exist between any two objects (Gold, 1991). It is therefore natural to store the

Fig. 2. The dual graph is used to describe the relationships between adjacent polygons.
398 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

Fig. 3. (a) The Delaunay triangulation for a set of points in the plane. (b) The Voronoi diagram for the same set
of points. (c) Both structures.

relationships between two given map objects in the edge that is dual to their shared bound-
ary (Fig. 2d).
The best known example of a graph and its dual is the Voronoi diagram and the Del-
aunay triangulation. In this case, geometric constraints are added so that the graphs
respect some criteria. Let S be a set of points in the Euclidean space R2 . Many triangula-
tions (a triangulation is a graph) of S are possible, but only one has the property that the
interior of the circumcircle of each triangle is empty of any other point in S: the Delaunay
triangulation, as shown in Fig. 3. The dual graph of the DT is the VD: each triangle in S
becomes a vertex in the dual (it is located at the centre of the circumcircle of the triangle)
and two vertices in the VD are linked by an edge if and only if their respective dual trian-
gles are adjacent. The resulting polygon dual to a point p 2 S is a Voronoi cell, denoted
Vp , which is defined by the set of points x 2 R2 that are closer to p than to any other point
in S. From a mathematical point of view:

Vp ¼ fx 2 R2 jkx  pk 6 kx  qk 8q 2 Sg:

The union of the Voronoi cells of all points p in S form VD(S). In any dimensions, a
Voronoi cell around a point p is always convex, and its size depends on the configuration
of neighbouring points. The VD and DT are very important structures in a wide range of
domains, see Aurenhammer (1991) for an excellent survey.
Both the VD and the DT generalise to any dimensions. The Voronoi cell Vp of a point
p in R3 has the same definition but becomes a polyhedron. The generalisation to 3D of the
Delaunay triangulation is the Delaunay tetrahedralization: each triangle becomes a tetra-
hedron that satisfies the empty circumsphere rule. The duality between the two structures in
3D is also similar: each tetrahedron of the DT becomes a vertex (located at the centre of
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 399

Fig. 4. Duality in 3D between the elements of the Voronoi diagram and the Delaunay tetrahedralization.

the circumsphere of the tetrahedron), and two vertices in the VD are linked by an edge if
and only if their two respective dual tetrahedra are adjacent.
More formally, the mapping between the elements of primal and dual subdivisions in
Rd is as follows. Let C be a k-cell of any structure, the dual cell of C in Rd is denoted
by Cw and is a (d  k)-cell. Therefore, in 3D, a point becomes a polyhedron, and vice
versa; and an edge becomes a face, and vice versa. For example, a Delaunay vertex p
becomes a Voronoi cell (Fig. 4a), a Delaunay edge a becomes a Voronoi face (Fig. 4b),
a Delaunay triangular face j becomes a Voronoi edge (Fig. 4c), and a Delaunay tetrahe-
dron s becomes a Voronoi vertex (Fig. 4d).

4. Augmented quad-edge

4.1. Quad-edge structure

The quad-edge data structure (Guibas & Stolfi, 1985) was developed to represent simul-
taneously the primal and dual subdivisions of a 2-manifold, and also to navigate from
edge to edge in these subdivisions. Each quad-edge represents one geometrical edge of a
subdivision, and this edge is decomposed into four directed edges: two in the primal,
and their two respective dual edges. Each directed edge is represented by a quad, which,
in a given subdivision, will either point to one vertex or to a face; in the dual, a quad points
to the dual of what it is pointing to in the primal. Its advantages are firstly that there is no
distinction between the primal and the dual representation (it is symmetric with respect to
edges and faces/polygons), and secondly that all operations are performed as pointer oper-
ations only, thus giving an algebraic representation to its operations. Fig. 5a shows the
basic structure, with four branches (quads) for each edge of the graph being stored (the
dual edges are not drawn), one for each of the incident vertices or faces. There are three
pointers on each quad: one to the vertex or face object (org), one to the next anticlockwise
quad-edge around that object (next), and one to link the four quads together in a loop
(rot). Thus all vertices or faces have complete topological loops around them. Notice that
the rot operation actually permits us to navigate between the primal and the dual
subdivisions.
Only two commands are used to modify a graph: MakeEdge to create a new edge on a
manifold, and Splice to connect/disconnect quad-edges together (illustrated in Fig. 5b). In
the simplest case, Splice connects two separate next loops, joining the two nodes together,
and at the same time splitting the next loop around the common face. Splice is its own
inverse.
400 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

Fig. 5. (a) The quad-edge data structure and some basic operators (only one subdivision is represented). The
starting quad q is the black quad, and the resulting quads are grey. (b) The Splice operator.

4.2. One dimension higher

The augmented quad-edge (AQE) uses the ‘normal’ quad-edge, which is valid for any
2-manifolds, to represent each 3-cell of a 3D complex, in either space. For instance, each
tetrahedron and each Voronoi cell are independently represented with the quad-edge,
which is akin to a boundary representation (b-rep). With this simple structure, it is possible
to navigate within a single cell with the quad-edge operators, but in order to do the same
for a 3D cell complex two things are missing: a way to ‘link’ adjacent cells in a given space,
and also a mechanism to navigate to the dual space. First, notice that in this case two of
the four org pointers of a quad-edge point to vertices forming the 2-manifold, but the
other two (which in 2D point to the dual, or a face) are not used in 3D. Remember also
that in 3D the dual of a face is an edge. Our idea is therefore to use this dual edge to ‘link’
two cells sharing a face: the unused face pointers simply point to their dual edge. This per-
mits us to ‘link’ cells together in either space, and also to navigate from a space to its dual.
Indeed, we may move from any quad-edge with a face pointer to a quad-edge in the dual
cell complex, and from there we may return to a different 2-manifold in the original cell
complex if needed.
A quad-edge is divided into four quads q, and there exist two types of quads: a qf points
to a face, and a qv points to a vertex. One rot operation applied to a qf returns a qv, and
vice-versa. A qf identifies uniquely, like the facet-edge (Dobkin & Laszlo, 1989), a pair
w
(face, edge). Therefore qf has also a linked quad qH f in the dual that is defined by (face ,
w
edge ).
One issue remains to be resolved: as each face is penetrated by several dual edges, a con-
sistent rule must be defined to select the appropriate one. Indeed, with the AQE, the dual
edge to a face has to be stored for all the dual cells sharing that edge. A triangular face has
for example three dual edges since each of its three vertices becomes a cell in the dual. A qv
has its org pointer set to a node, and a qf has its org pointer set to qH H
f . The pointer to qf
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 401

Fig. 6. The through pointer. The quad qf is the starting quad, and the grey ones are resulting quads after an
operation.

from qf is called through, as shown in Fig. 6. The quad qH f is defined as belonging to the
dual cell which encloses the node pointed to by qf.rot = qv. This is sufficient to define
the through pointer structure. In Fig. 6, the triangular face abc has three dual edges and
may be represented by the quad qf (the black quad) on the edge ab. The dual edge of
abc is uv, and the grey face represents the face that is dual to the edge ab. The operation
qf.through gives qHf which is situated on the edge uv and on the face dual to ab; this face
belongs to the cell dual to vertex b since qf.rot.org gives b. Note also that qf.next.through
gives a quad on a face which belongs to the cell dual to the vertex a; and qf.next.next.
through to c.

4.3. Navigation and construction operators

As mentioned earlier, because the quad-edge structure is used to represent each cell of a
3D cell complex, the ‘usual’ navigation operators (next and rot) described in Section 4.1
can be used to navigate within a 3-cell. Only one new operator is needed to navigate within
a 3D cell complex and to be able to visit every polyhedron, face, node or quad-edge: the
through operator, as described in the previous section.
Let qf be a quad and qHf its dual quad, then the following properties are valid:

1. qH
f ¼ qf :through
2. qf ¼ qf :through:through ¼ qH
f :through

Different higher-level navigation operators can also be defined. One that is particularly
useful is an operator for ‘jumping’ from a cell to another adjacent cell, in either space. We
call this navigation operator adjacent. Let qf be a quad on the boundary of a common face
between two cells, as shown in Fig. 7. We want to ‘jump’ to the quad rf representing the
same pair (face, edge) as qf but on the adjacent cell. We do this by using the dual structure
to navigate. Starting from qf, qf.through gives qH H
f , and qf :next gives a quad whose through
pointer gives a quad in the original space but in a neighbouring cell. The rot operator is
used twice to set the result to rf. Thus,
402 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

Fig. 7. The adjacent operator. The quads qf and rf represent the same pair (face, edge), but belong to different
3-cells.

qf :adjacent ¼ qf :through:next:through:rot2
Other relations are also interesting. First, the adjacent operator is symmetrical:
qf = qf.adjacent.adjacent. Second, if qf and rf are adjacent quads as in Fig. 7 (such that
rf = qf.adjacent), notice that their respective dual quads qH H
f and r f are also adjacent, i.e.,
they represent the same pair (face, edge) but are part of different 3-cells.
New construction operators are required to build a 3D cell complex and its dual. The
creation of new independent cells, e.g. a Delaunay tetrahedron or a Voronoi cell, can be
done with only the construction operators MakeEdge and Splice of the quad-edge struc-
ture because each cell is represented as a 2-manifold. The other construction operator
needed is for the assemblage or linkage of two cells. Again, this is done fairly easily as only
the through pointers of all qf belonging to a common face need to be set (reset) to link (un-
link) two cells.

4.4. Properties

We have attempted to develop a data structure with a variety of properties. The AQE
integrates the b-rep representations and the irregular and regular decompositions under
the same structure, as described in Section 2. Its flexibility, and the fact that the dual of
a 3D cell complex is also stored, makes the AQE a valuable choice for many applications,
as we discuss in Section 5. It should be noticed that the AQE is only valid for cell com-
plexes, and that subdivisions where ‘dangling’ elements are present cannot be represented
with the structure in its current form.
The proposed approach preserves the lower dimensionality (2D) navigation and con-
struction operations, as well as keeping the underlying quad-edge properties in 3D. As
a consequence, the data structure is directly navigable and forms an ‘edge algebra’, allow-
ing us to combine our pointer operations by following well defined rules, and to show the
equivalence of different ways of getting to the same destination. Also, every operation has
an inverse. To permit the navigation between 2-manifolds, the dual has to be explicitly
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 403

stored within the structure. Each volumetric element uses the quad-edge (with 2D pri-
mal and dual edge elements) for navigation around its boundary, and the same is thus
required for the dual structure, as we wish the same set of operations to be valid in each
space.
Furthermore, the AQE has the major benefit of being locally modifiable, i.e., if new ele-
ments need to be inserted (deleted) in (from) a model a complete reconstruction of the
model is not required. We have already implemented with the AQE an algorithm to con-
struct and store both the 3D DT and the VD. It is based on local transformations, called
flips, to modify a tetrahedralization (Joe, 1991). A DT is constructed incrementally by
inserting one at a time the points in a set and updating the DT – with flips – after each
insertion. Our modifications involve the update of the dual structure when a flip is per-
formed; because the flips are simple topological operations involving a fixed set of tetra-
hedra, the operations to update the VD are relatively simple and can be described using
algebraic statements exclusively. This simplicity also means that vertices can be deleted
from a 3D VD/DT (Ledoux, Gold, & Baciu, 2005), which is of great value for many appli-
cations. More details concerning the construction and the modification of the 3D VD/DT
with the AQE will be published in another paper.

4.5. Storage costs and comparisons

We analyse in this section the storage costs of the AQE when it is used to represent both
the DT and the VD.
With the AQE, each cell (tetrahedron or Voronoi cell) is stored separately. Each tetra-
hedron contains: six edges, each represented by four quads containing three pointers (org,
next and rot). This makes a total of 72 pointers. Notice that every qf has one and only one
qHf , and also that the same is true for every qv (there is always one qv when a rot is applied
to a given qf). As a result, the total number of pointers for the dual is also 72, which makes
a total of 144 pointers for each tetrahedron.
This is obviously more than the data structures used to store only the DT. For example,
the tetrahedron-based structure described in Section 2 uses only eight pointers per tetra-
hedron. To our knowledge, the only data structure preserving both a primal and dual
3D subdivision is the facet-edge (Dobkin & Laszlo, 1989). With it, each pair (face, edge)
contains (considering both the primal and the dual): four pointers to vertices (two Dela-
unay and two Voronoi), and also four pointers to the next facet-edge (the navigation
requires the use of the dual to access the adjacent facet-edges around the face and the
edge). There are three pairs per triangular face, and a tetrahedron contains four faces.
However, with this structure, the common faces between two tetrahedra are not stored
twice – and we can state that each face is shared by two tetrahedra, except the ones on
the convex hull of the data set. As a result, the total number of pointers per tetrahedron
is 48, but, as explained previously, extra information needs to be stored if one wants to use
the quad-edge operators to navigate within a single cell.
To summarise, the AQE, as described in Section 4, is about three times more space-con-
suming than the facet-edge, in the worst case. A factor of two is consistent with our initial
objective to maintain the primal and dual structures simultaneously, and to use the same
set of operations in each space. This can be achieved by compressing the quad-edge as
Guibas and Stolfi proposed: the rot pointers between quads are not explicitly stored
and the last two bits of each pointer are used to identify each quad of a quad-edge object.
404 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

5. Applications

Many applications in 3D are based on the idea of subdividing or tessellating the space
to be able to model objects. This section gives a few examples of applications for which the
dual of a subdivision is meaningful, and therefore is worth storing. Doing this can opti-
mise some of the operations necessary for an application, solve a problem that data struc-
tures storing only one subdivision have, or offer a more flexible option for the storage of
attributes.

5.1. Crust and skeleton

In recent years, tetrahedralizations have been used more and more to model the interior
and the boundary of objects. The principal application is in engineering where the subdivi-
sion of an object is used in the finite element method or other numerical methods for solv-
ing partial differential equations. Delaunay tetrahedra are usually preferred because of
their properties, but are sometimes refined so that they respect some criteria (see Shewchuk
(1997) for a discussion). This is mostly the reason why the efforts of the computational
geometry community have been directed towards data structures to efficiently store tetra-
hedra, and only tetrahedra.
But some applications are only possible when both the DT and the VD are known
explicitly, i.e., that properties of the DT and the VD are both used to solve a problem.
Obvious two-dimensional examples are the definitions of the crust and skeleton of a suf-
ficiently well-sampled polygon. It has been shown that these two structures are subsets of
respectively the DT and the VD (Amenta, Marshall, & Eppstein, 1998), and the extraction
of either the crust or the skeleton requires the knowledge of both the DT and the VD.
Gold and Snoeyink (2001) show that the crust/skeleton concept helps in many carto-
graphic problems such as the reconstruction of terrain models from contour lines, the esti-
mation of a river network from boundaries, and the extraction of ‘‘topology’’ from
scanned maps. One interesting application is polygon generalisation: a polygon may be
generalised by first constructing its skeleton, then simplifying this skeleton, and finally
reconstructing the polygon from the simplified skeleton. This idea has been extended to
three dimensions for the reconstruction and simplification of surfaces and features from
laser scans (Tam & Heidrich, 2003). The method is similar to the two-dimensional one
and involves the extraction of the DT of the set of points, and the 3D skeleton of the
reconstructed objects (with the VD). The surface is simplified by modifying the skeleton,
which as a consequence involves modifying the crust.

5.2. Modelling 3D datasets in the geosciences

In the same way that triangulated irregular networks are used in two dimensions to rep-
resent the spatial variation of an attribute (usually the elevation of a terrain), tetrahedra
can be used to model geoscientific datasets, as an alternative to raster structures (Lattuada,
1998). Examples are datasets in geology or oceanography where continuous phenomena
(e.g. the percentage of gold in the rock or the salinity of the water) are modelled. However,
as Ledoux (2006) explain, modelling 3D continuous phenomena can also benefit greatly
from both the DT and the VD. As discussed in Section 2, most of the algorithms and data
structures available for constructing and storing the VD actually use only tetrahedra and
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 405

extract the VD when needed. Although this results in a faster implementation, it has one
major drawback if one wants to work with the VD: the computation of the Voronoi ele-
ments (vertices, edges, faces and polyhedra) is a computationally expensive operation.
Although a single Voronoi vertex is easily extracted (it is located at the centre of the tet-
rahedron to which it is dual), the extraction of the other elements is more complex. The
next examples all refer to Fig. 4. A Voronoi edge, which is dual to a triangular face j,
is formed by the two Voronoi vertices dual to the two tetrahedra sharing j. A Voronoi
face, which is dual to a Delaunay edge a, is formed by all the vertices that are dual to
the Delaunay tetrahedra incident to a. The idea is simply to ‘‘turn’’ around a Delaunay
edge and extract all the Voronoi vertices to extract a face. The construction of one Voro-
noi cell Vp dual to a vertex p is similar: it is formed by all the Voronoi vertices dual to the
tetrahedra incident to p. Since a Voronoi cell is convex by definition, it is possible to collect
all the Voronoi vertices and then compute the convex hull (see Barber, Dobkin, & Huh-
danpaa (1996) for more details). A simpler method consists of first identifying all the edges
incident to p, and then extracting the dual face of each.
In short, in order to model a geoscientific dataset, the DT is an important concept
because it helps in the construction and the manipulation of the spatial relationships
between the elements, but also because many 3D visualization algorithms are optimised
when tetrahedra are involved (Cignoni, Montani, & Scopigno, 1998). On the other hand,
the VD can be of great help with many spatial analysis operations, and storing it explicitly
considerably optimises the operations. Examples are interpolation methods (e.g. the nat-
ural neighbour interpolation method of Sibson (1981)), spatial analysis functions (Okabe,
Boots, & Sugihara, 1994), or traditional GIS operations (Gold, Remmele, & Roos, 1997).

5.3. Modelling apartment buildings

Perhaps the main advantage of using the AQE is that attributes can be stored for any
elements of both subdivisions (vertices, edges, faces and volumes). Indeed, the vertex attri-
butes may be stored along with the x  y  z coordinates in the vertex; edge attributes in a
quad pointing to a vertex; face attributes in a quad pointing to a face; and volume attri-
butes in the vertex that is dual to a volume.
One potential three-dimensional application exploiting duality to store attributes is the
modelling of an apartment building. Each room is represented by one 3-cell (see Fig. 8
for a simple example), and in a similar way to the example in Fig. 2, the dual edge to
each face (e.g. the four walls, the floor or the ceiling) is used to store the relationships

Fig. 8. An apartment building can be modelled with the AQE.

406 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

between two adjacent rooms, such as the visibility (instead of explicitly representing a
window for instance), or the noise transmission. Moreover, it is possible to store attri-
butes for the properties of different parts of one room: the finish of a wall or the floor
type (attributes attached to 2-cells), the type of moulding (attributes attached to the
1-cells), or the properties for the whole room (e.g. its volume or renting price) can be
stored in the dual vertex to the 3-cell. A major advantage of using the AQE for modelling
apartment buildings is that a wall (or a floor) is stored twice, once for each of the two
rooms sharing the wall. This is consistent with the way one would conceptualise many
rooms in a building: the wall separating two rooms will not necessarily have the same
properties on each side. For example, the finish on one side of the wall might be different
from the one on the other side, or one side might have a gold ceiling moulding and the
other no moulding at all.
The AQE can also be used to search efficiently a 3D cell complex. Referring to Fig. 8, if
qf (the black quad) is a quad representing the wall between the two rooms, it is possible to
retrieve all the elements forming the room on the right by using a simple graph search
algorithm such breadth-first search (BFS) on the 3-cell representing the room (using
quad-edge operators). Each element of the wall that qf represents can also be retrieved
with the quad-edge operators rot and next; and the same is true for the ceiling or the floor
that can be accessed directly from the AQE algebra, without searching. The attributes cha-
racterising the relation between the two rooms are obtained with qf.through, and then the
attributes of the whole room can be obtained directly because they are stored in the vertex
dual to the room. Also, from qf it is possible to navigate locally and visit the adjacent room
with qf.adjacent, and then for instance move to its ceiling by applying rot twice. Given a k-
cell C, it is actually possible to identify all the l-cells (where k < l 6 3) incident to C using
the navigation operators of Section 4.3. Notice that in the case of retrieving all the ele-
ments incident to a vertex one can simply move to the dual subdivision and perform a
BFS on the 3-cell because the elements retrieved are dual to elements incident in the pri-
mal. A further important consideration is that a complete 3D cell complex can be easily
and efficiently traversed without storing extra information (de Berg, van Kreveld, van
Oostrum, & Overmars, 1997).
When the attributes are included in a search, it is for example possible to look for an
edge that represents gold ceiling moulding, incident to a white ceiling and a blue wall,
and with a red corner boss; or for all the rooms having more than 100 m2 but not having
a carpet.

6. Discussion

This paper has outlined a new data structure to store simultaneously a 3D cell complex
and its dual, and briefly showed its use for different applications. The augmented quad-
edge integrates under one structure different types of data structures used in different
application domains. While the data structure requires somewhat more storage than some
more simple structures, it has the advantage of conceptual simplicity and involves only a
simple extension of the 2D topological relationships. The fact that the dual subdivision is
also stored permits the storage of attributes for any elements of a 3D subdivision, and
makes the structure particularly valuable for a wide range of applications. We believe that
the AQE offers a good trade-off between storage costs and the number of topological rela-
tionships between the different elements of a subdivision.
 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408 407

We are still exploring if the structure is appropriate for arbitrary subdivisions of

3-manifolds where ‘dangling’ faces and edges may exist. Future work also includes more
formal definition of navigation and construction operations, and possible compression of
the structure. Earlier work suggests that quad-edges, being an algebra, may be stored effec-
tively using relational databases, but this is still being examined (Merrett, 2005).

Acknowledgements

Part of this work was done when the authors were at the Hong Kong Polytechnic
University, and they thank the Hong Kong’s Research Grants Council (Project PolyU
5068/00E) for its support. They would also like to thank Marcin Dzieszko for helping with
the programming. Finally, the second author acknowledges the funding of the EU Marie
Curie Chair.

References

Amenta, N., Marshall, B., & Eppstein, D. (1998). The crust and the beta-skeleton: combinatorial curve
reconstruction. Graphical Models and Image Processing, 60(2), 125–135.
Anselin, L. (1999). Interactive techniques and exploratory spatial data analysis. In P. A. Longley, M. F.
Goodchild, D. J. Maguire, & D. W. Rhind (Eds.), Geographical information systems (pp. 253–266). John Wiley
& Sons.
Aurenhammer, F. (1991). Voronoi diagrams: A survey of a fundamental geometric data structure. ACM
Computing Surveys, 23(3), 345–405.
Barber, C. B., Dobkin, D. P., & Huhdanpaa, H. T. (1996). The Quickhull algorithm for convex hulls. ACM
Transactions on Mathematical Software, 22(4), 469–483.
Baumgart, B. G. (1975). A polyhedron representation for computer vision. In National computer conference.
AFIPS.
Bertrand, Y., Dufourd, J. F., Françon, J., & Lienhardt, P. (1993). Algebraic specification and development in
geometric modeling. In Proceedings of TAPSOFT’93, Orsay, France. LNCS (Vol. 668, pp. 75–89).
Boissonnat, J.-D., Devillers, O., Pion, S., Teillaud, M., & Yvinec, M. (2002). Triangulations in CGAL.
Computational Geometry – Theory and Applications, 22, 5–19.
Boudriault, G. (1987). Topology in the TIGER file. In Proceedings 8th international symposium on computer
assisted cartography. Baltimore, USA.
Cignoni, P., Montani, C., & Scopigno, R. (1998). Tetrahedra based volume visualization. In H.-C. Hege & K.
Polthier (Eds.), Mathematical Visualization – Algorithms, Applications, and Numerics (pp. 3–18). Springer-
Verlag.
de Berg, M., van Kreveld, M., van Oostrum, R., & Overmars, M. (1997). Simple traversal of a subdivision
without extra storage. International Journal of Geographical Information Science, 11(4), 359–374.
De Floriani, L. & Hui, A. (2003). A scalable data structure for three-dimensional non-manifold objects. In
Proceedings 1st Eurographics symposium on geometry processing (pp. 72–82). Aachen, Germany.
Dobkin, D. P., & Laszlo, M. J. (1989). Primitives for the manipulation of three-dimensional subdivisions.
Algorithmica, 4, 3–32.
Gold, C. M. (1991). Problems with handling spatial data – the Voronoi approach. CISM Journal, 45(1), 65–80.
Gold, C.M. (1993). Forestry spatial decision support system classification, and the ‘flight simulator’ approach. In
Proceedings GIS’93: Eyes on the future (pp. 797–802). Vancouver, Canada.
Gold, C. M., Remmele, P. R., & Roos, T. (1997). Voronoi methods in GIS. In M. van Kreveld, J. Nievergelt, T.
Roos, & P. Widmayer (Eds.), Algorithmic foundations of geographic information systems. Lecture notes in
computer science (1340, pp. 21–35). Springer-Verlag.
Gold, C. M., & Snoeyink, J. (2001). A one-step and skeleton extraction algorithm. Algorithmica, 30, 144–163.
Guibas, L. J., & Stolfi, J. (1985). Primitives for the manipulation of general subdivisions and the computation of
Voronoi diagrams. ACM Transactions on Graphics, 4, 74–123.
Joe, B. (1991). Construction of three-dimensional Delaunay triangulations using local transformations. Computer
Aided Geometric Design, 8, 123–142.
408 H. Ledoux, C.M. Gold / Comput., Environ. and Urban Systems 31 (2007) 393–408

Lattuada, R. (1998). A triangulation based approach to three dimensional geoscientific modelling. Ph.D. thesis,
Department of Geography, Birkbeck College, University of London, London, UK.
Ledoux, H. (2006). Modelling three-dimensional fields in geoscience with the Voronoi diagram and its dual. Ph.D.
thesis, School of Computing, University of Glamorgan.
Ledoux, H., Gold, C. M., & Baciu, G. (2005). Flipping to robustly delete a vertex in a Delaunay
tetrahedralization. In Proceedings international conference on computational science and its applications –
ICCSA 2005. LNCS (3480, pp. 737–747). Singapore: Springer-Verlag.
Lienhardt, P. (1994). N-dimensional generalized combinatorial maps and cellular quasi-manifolds. International
Journal of Computational Geometry and Applications, 4(3), 275–324.
Lopes, H. & Tavares, G. (1997). Structural operators for modeling 3-manifolds. In Proceedings 4th ACM
symposium on solid modeling and applications (pp. 10–18). Atlanta, Georgia, USA.
Mäntylä, M. (1988). An introduction to solid modeling. New York, USA: Computer Science Press.
Merrett, T. H. (2005). Topological structures for spatial databases in two and three dimensions. In Proceedings
17th Australasian database conference (ADC2005). Newcastle, Australia.
Morehouse, S. (1985). ARC/INFO: A geo-relational model for spatial information. In Proceedings 7th
international symposium on computer assisted cartography. Washington DC, USA.
Muller, D. E., & Preparata, F. P. (1978). Finding the intersection of two convex polyhedra. Theoretical Computer
Science, 7, 217–236.
Okabe, A., Boots, B., & Sugihara, K. (1994). Nearest neighbourhood operations with generalized Voronoi
diagrams: a review. International Journal of Geographical Information Systems, 8(1), 43–71.
Requicha, A. A. G. (1982). Representation of rigid solids – theory, methods and systems. ACM Computing
Surveys, 12(4), 437–464.
Samet, H. (1984). The quadtree and related hierarchical data structures. ACM Computing Surveys, 16(2),
187–260.
Shewchuk, J. R. (1997). Delaunay Refinement Mesh Generation. Ph.D. thesis, School of Computer Science,
Carnegie Mellon University, Pittsburg, USA.
Sibson, R. (1981). A brief description of natural neighbour interpolation. In V. Barnett (Ed.), Interpreting
multivariate data (pp. 21–36). New York, USA: Wiley.
Tam, R., & Heidrich, W. (2003). Shape simplification based on the medial axis transform. In Proceedings IEEE
visualization 2003. Seattle, Washington, USA: IEEE Computer Society.
Zlatanova, S., Abdul Rahman, A., & Shi, W. (2004). Topological models and frameworks for 3D spatial objects.
Computers and Geosciences, 30, 419–428.